Key Takeaways
- Google has released a new episode of its official podcast, “Made by Google Podcasts,” titled “Smart Silicon: Tensor G5 and the Next Era of the AI Phone | Made by Google Podcast S8E4.”
- The episode delves into the “Tensor G5,” the original processor debuting in the next generation of Google Pixel devices: “Pixel 10,” “Pixel 10 Pro,” “Pixel 10 Pro XL,” and “Pixel 10 Pro Fold.”
- Key points discussed include the processor’s ability to handle increasingly large AI models, its new technologies for performance and efficiency, and a development focus on practical usability.
On Thursday, September 25, 2025, Google released a new episode of its official podcast, “Made by Google Podcasts.” Titled “Smart Silicon: Tensor G5 and the Next Era of the AI Phone | Made by Google Podcast S8E4,” the episode explores the original processor, “Tensor G5,” which will be featured in the next generation of Google Pixel devices: “Pixel 10,” “Pixel 10 Pro,” “Pixel 10 Pro XL,” and “Pixel 10 Pro Fold.”
In “Smart Silicon: Tensor G5 and the Next Era of the AI Phone | Made by Google Podcast S8E4,” Jesse Seed, Group Product Manager for the Google Silicon team, discusses how “Tensor G5” will revolutionize AI capabilities in the upcoming “Pixel 10” series. The key discussion points revolve around three main areas: the ability to handle larger AI models, new technologies for performance and efficiency, and a development approach prioritizing practical use.
Handling Increasingly Large AI Models
AI models in smart devices are rapidly becoming more complex and, consequently, larger.
For instance, the number of parameters in AI models used for camera zoom functions has increased from tens of thousands in 2021 to millions today. The “Tensor G5” is capable of directly running models with nearly a billion parameters on-device for features like “ProRes Zoom.”
To efficiently run these increasingly large AI models within the limited power and thermal constraints of Google Pixel devices, processors like the custom-designed “Tensor G5” are essential.
New Technologies for Performance and Efficiency
The “Tensor G5” features advancements through two new technologies: 3nm process technology and the “Matformer” architecture.
The 3nm process technology is a micro-fabrication technique that allows for more features and processing power to be packed into the same area compared to other processors. This results in significant improvements in both performance and power efficiency.
The “Matformer” architecture is an AI model structure inspired by Russian nesting dolls, or “Matryoshka.” It incorporates smaller, faster AI sub-models within larger AI models designed for maximum quality.
The “Tensor G5” dynamically switches between these AI models, enabling it to achieve both high performance and efficiency. For example, in the Google Pixel-exclusive “Recorder” app, processing speed has increased by 2.6 times and energy efficiency has improved by 2 times compared to the previous generation (Tensor G4).
Development Focused on Practical Usability
Google prioritized improvements that users can actually experience over benchmark scores when developing the “Tensor G5.”
As a result, the TPU computing power dedicated to AI processing has increased by up to 60%, and CPU performance has seen an average improvement of 34%. These advancements were made to ensure the smooth operation of AI features for the new generation of Google Pixel devices, such as “ProRes Zoom” and “Smart Suggest.”
These improvements are positioning the new generation of Google Pixel devices, the “Pixel 10” series, to transition from being merely high-spec smartphones to true “AI Phones.”
Source: Google
コメントを残す